Fix CUDSS extension for v0.5 and v0.6 compatibility

Updates the ArrayInterfaceCUDSSExt to use the modern CUDSS API.
The deprecated CudssMatrix type has been replaced with standard
CUDA.CuVector types, which is the current API pattern in CUDSS v0.5+.

Changes:
- Replace CudssMatrix with CUDA.CuVector in lu_instance
- Add CUDA as explicit extension dependency
- Add CUDSS v0.5 and v0.6 to compatibility bounds
- Bump version to 7.20.1

Fixes the runtime incompatibility introduced by the breaking changes
in CUDSS v0.5 and v0.6 where CudssMatrix was removed in favor of
standard CUDA.jl types (CuVector).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: ChrisRackauckas

Project.toml

@@ -1,6 +1,6 @@
 name = "ArrayInterface"
 uuid = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
-version = "7.20.0"
+version = "7.20.1"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
@@ -24,7 +24,7 @@ Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
 ArrayInterfaceBandedMatricesExt = "BandedMatrices"
 ArrayInterfaceBlockBandedMatricesExt = "BlockBandedMatrices"
 ArrayInterfaceCUDAExt = "CUDA"
-ArrayInterfaceCUDSSExt = "CUDSS"
+ArrayInterfaceCUDSSExt = ["CUDSS", "CUDA"]
 ArrayInterfaceChainRulesCoreExt = "ChainRulesCore"
 ArrayInterfaceChainRulesExt = "ChainRules"
 ArrayInterfaceGPUArraysCoreExt = "GPUArraysCore"
@@ -39,7 +39,7 @@ Adapt = "4"
 BandedMatrices = "1"
 BlockBandedMatrices = "0.13"
 CUDA = "5"
-CUDSS = "0.2, 0.3, 0.4"
+CUDSS = "0.2, 0.3, 0.4, 0.5, 0.6"
 ChainRules = "1"
 ChainRulesCore = "1"
 ChainRulesTestUtils = "1"

ext/ArrayInterfaceCUDSSExt.jl

@@ -2,14 +2,18 @@ module ArrayInterfaceCUDSSExt
 
 using ArrayInterface
 using CUDSS
+using CUDA
 
 function ArrayInterface.lu_instance(A::CUDSS.CuSparseMatrixCSR)
     ArrayInterface.LinearAlgebra.checksquare(A)
-    fact = CudssSolver(A, "G", 'F')
     T = eltype(A)
-    n = size(A,1)
-    x = CudssMatrix(T, n)
-    b = CudssMatrix(T, n)
+    n = size(A, 1)
+
+    # Use standard CUDA types (CuVector) instead of deprecated CudssMatrix
+    x = CUDA.CuVector{T}(undef, n)
+    b = CUDA.CuVector{T}(undef, n)
+
+    fact = CudssSolver(A, "G", 'F')
     cudss("analysis", fact, x, b)
     fact
 end